refining-hypothesis

Refining Hypothesis

Takes a hypothesis stub and expands it through targeted literature research. Searches for supporting evidence, counterexamples, and related work, then sharpens the hypothesis into a precise, research-grounded claim ready for experimentation.

The orchestrator runs this skill in parallel across all hypotheses in the current iteration. Each invocation is independent and works on a single hypothesis file.

---

Arguments

• Hypothesis file -- path to hypothesis-NN.md

---

Workflow

Step 0: Check state (idempotent checkpoint)

Read the hypothesis file. If it already contains a ## Literature section, this hypothesis has already been refined -- report "Already refined: " and stop. The file itself is the checkpoint, so re-running is safe.

Identify the problem directory from the file path (its parent directory). Read problem.md to understand the broader research context. Read references.md if it exists to know which sources are already in the bibliography -- this prevents the literature agents from re-searching known material.

Step 1: Build search queries

Derive 3 targeted search queries from the hypothesis statement and rationale. Each query targets a different angle because a single search would only surface one facet of the literature:

1. The claim itself -- look for direct confirmation or refutation of this specific hypothesis
2. The underlying mechanism -- what would need to be true for this hypothesis to hold? Search for evidence about that prerequisite
3. Known counterexamples -- cases where similar claims have failed or been disproven

Step 2: Run literature research in parallel

Spawn 3 Task agents in a single response (one per query). Each Task runs /researching-literature targeting the problem directory.

Task prompt format for each:

Run /researching-literature with:
- Topic: "<search query>"
- Problem directory: "<absolute path to problem directory>"

Wait for all three to complete. Then read the updated references.md to identify which new REF IDs were added. Note the relevant ones for this hypothesis.

Step 2b: Gap assessment and full dual-track search

Read the updated references.md. Assess coverage along two dimensions:

• Recency gap: are any sources from the last 3 years present? If all found sources are older than 3 years, recent work is underrepresented.
• Landmark gap: are any foundational or definitional sources present? If all found sources are very recent, foundational context is missing.

Run one full /researching-literature invocation to fill gaps. Always run this step — the three parallel searches in Step 2 are targeted and narrow; this invocation provides broad dual-track coverage.

Task prompt format:

Run /researching-literature with:
- Topic: "<broader query derived from the hypothesis topic, not the specific claim>"
- Problem directory: "<absolute path to problem directory>"

Wait for the invocation to complete, then read references.md again to identify additional REF IDs added.

Step 3: Sharpen the hypothesis

With the literature findings in hand, assess the hypothesis through these lenses:

• Supporting sources -- which REF IDs provide evidence for the claim?
• Challenging sources -- which REF IDs complicate, contradict, or narrow the claim?
• Adjustment needed? -- does the original statement need narrowing (too broad), broadening (too narrow), or reframing (wrong angle)?
• Already resolved? -- if the literature definitively confirms or refutes the hypothesis through prior work, note this clearly. It changes the downstream pipeline: confirmed/refuted hypotheses skip experiment design.

Step 4: Update hypothesis file

Append to the hypothesis file after the existing content. Preserve the original ## Statement and ## Rationale sections unchanged -- they serve as a historical record of the initial thinking.

Use Edit to append the following sections:

Literature section:

## Literature

Relevant sources: REF-NNN, REF-NNN, ...

### Supporting evidence
- <finding> (REF-NNN)
- <finding> (REF-NNN)

### Challenges and counterexamples
- <finding that complicates or contradicts this hypothesis> (REF-NNN)

Refined statement section:

## Refined Statement
<Updated hypothesis if the original needed adjustment. If unchanged, write "Unchanged -- original statement holds.">

If the hypothesis is already resolved by literature (definitively confirmed or refuted by prior work), add a note explaining this and update ## Status from pending to confirmed or refuted accordingly. Use Edit to change the status line.